diff --git a/include/rabitqlib/clustering/qgkmeans.hpp b/include/rabitqlib/clustering/qgkmeans.hpp
index 5989692f..d008d759 100644
--- a/include/rabitqlib/clustering/qgkmeans.hpp
+++ b/include/rabitqlib/clustering/qgkmeans.hpp
@@ -184,25 +184,7 @@ class QGAssigner {
     void clear_hints() { previous_labels_.clear(); }
 
    private:
-    void assign_with_cache(
-        const float* centroids,
-        const float* x,
-        size_t n,
-        PID* labels,
-        float* distances,
-        detail::QuerySumCache& query_sums
-    ) {
-        if (centroids == nullptr || x == nullptr || labels == nullptr ||
-            distances == nullptr) {
-            throw std::invalid_argument(
-                "QGAssigner input and output pointers must not be null"
-            );
-        }
-
-        if (n > static_cast<size_t>(std::numeric_limits<std::ptrdiff_t>::max()) / d_) {
-            throw std::invalid_argument("QGAssigner input dimensions overflow ptrdiff_t");
-        }
-
+    void rebuild_centroid_graph(const float* centroids) {
         if (builder_) {
             builder_->reset(centroids);
         } else {
@@ -227,6 +209,28 @@ class QGAssigner {
         } else {
             builder_->build(niter_);
         }
+    }
+
+    void assign_with_cache(
+        const float* centroids,
+        const float* x,
+        size_t n,
+        PID* labels,
+        float* distances,
+        detail::QuerySumCache& query_sums
+    ) {
+        if (centroids == nullptr || x == nullptr || labels == nullptr ||
+            distances == nullptr) {
+            throw std::invalid_argument(
+                "QGAssigner input and output pointers must not be null"
+            );
+        }
+
+        if (n > static_cast<size_t>(std::numeric_limits<std::ptrdiff_t>::max()) / d_) {
+            throw std::invalid_argument("QGAssigner input dimensions overflow ptrdiff_t");
+        }
+
+        rebuild_centroid_graph(centroids);
         const auto distance_func = graph_->metric_type() == METRIC_IP
                                        ? dot_product_dis<float>
                                        : euclidean_sqr<float>;
diff --git a/python_bindings/bindings_common.hpp b/python_bindings/bindings_common.hpp
index e9d05cd0..85dc1783 100644
--- a/python_bindings/bindings_common.hpp
+++ b/python_bindings/bindings_common.hpp
@@ -225,6 +225,26 @@ ensure_1d_integer_array(py::handle value, const char* name) {
     return ensure_1d_array<int64_t>(value, name);
 }
 
+// The caller has already validated rank/dtype. Removed points still occupy IDs.
+inline std::vector<PID> checked_removal_ids(
+    const py::array_t<int64_t, py::array::c_style | py::array::forcecast>& ids,
+    size_t num_stored_points
+) {
+    std::vector<PID> result;
+    result.reserve(static_cast<size_t>(ids.shape(0)));
+    const rabitqlib::detail::ConstPackedArrayView<int64_t> values(
+        reinterpret_cast<const char*>(ids.data())
+    );
+    for (py::ssize_t i = 0; i < ids.shape(0); ++i) {
+        const int64_t id = values[static_cast<size_t>(i)];
+        if (id < 0 || static_cast<uint64_t>(id) >= num_stored_points) {
+            throw std::invalid_argument("ids must be in [0, num_stored_points)");
+        }
+        result.push_back(static_cast<PID>(id));
+    }
+    return result;
+}
+
 // Validate before narrowing, and own the result for APIs that mutate assignments.
 inline std::vector<PID> checked_cluster_ids(
     py::handle value,
diff --git a/python_bindings/hnsw_bindings.cpp b/python_bindings/hnsw_bindings.cpp
index 318b97ae..4d38191a 100644
--- a/python_bindings/hnsw_bindings.cpp
+++ b/python_bindings/hnsw_bindings.cpp
@@ -7,7 +7,6 @@
 
 #include <algorithm>
 #include <cstddef>
-#include <cstdint>
 #include <limits>
 #include <memory>
 #include <stdexcept>
@@ -18,7 +17,6 @@
 #include "bindings_common.hpp"
 #include "rabitqlib/index/hnsw/hnsw.hpp"
 #include "rabitqlib/index/types.hpp"
-#include "rabitqlib/utils/packed.hpp"
 
 namespace py = pybind11;
 
@@ -156,20 +154,7 @@ class HnswIndex {
         if (!built_) {
             throw std::runtime_error("HnswIndex must be built or loaded before remove");
         }
-        // Removed labels remain allocated; validate against the full stored ID range.
-        const size_t count = index_->num_stored_points();
-        std::vector<rabitqlib::PID> to_remove;
-        to_remove.reserve(static_cast<size_t>(ids_array.shape(0)));
-        const rabitqlib::detail::ConstPackedArrayView<int64_t> values(
-            reinterpret_cast<const char*>(ids_array.data())
-        );
-        for (py::ssize_t i = 0; i < ids_array.shape(0); ++i) {
-            const int64_t id = values[static_cast<size_t>(i)];
-            if (id < 0 || static_cast<uint64_t>(id) >= count) {
-                throw std::invalid_argument("ids must be in [0, num_stored_points)");
-            }
-            to_remove.push_back(static_cast<rabitqlib::PID>(id));
-        }
+        auto to_remove = checked_removal_ids(ids_array, index_->num_stored_points());
         py::gil_scoped_release release;
         return index_->remove(to_remove.data(), to_remove.size());
     }
@@ -314,8 +299,8 @@ class HnswIndex {
 void register_hnsw(py::module_& m) {
     using namespace rabitqlib::python_bindings;
 
-    py::class_<HnswIndex>(
-        m, "HnswIndex", R"doc(An approximate nearest-neighbor index using HNSW.
+    static constexpr char kIndexDoc[] =
+        R"doc(An approximate nearest-neighbor index using HNSW.
 
 Parameters
 ----------
@@ -338,8 +323,9 @@ random_seed : int, optional
 Notes
 -----
 Build or load the index before searching or saving. Concurrent reads are
-supported; an overlapping mutation raises RuntimeError.)doc"
-    )
+supported; an overlapping mutation raises RuntimeError.)doc";
+
+    py::class_<HnswIndex>(m, "HnswIndex", kIndexDoc)
         .def(
             py::init<size_t, size_t, size_t, size_t, size_t, const std::string&, size_t>(),
             py::arg("dim"),
@@ -349,30 +335,7 @@ supported; an overlapping mutation raises RuntimeError.)doc"
             py::arg("nbits") = 8,
             py::arg("metric") = "l2",
             py::arg("random_seed") = 100,
-            R"doc(An approximate nearest-neighbor index using HNSW.
-
-Parameters
-----------
-dim : int
-    Original vector dimension, from 64 through 65536.
-max_elements : int
-    Storage capacity, including removed points. Use resize to grow it.
-M : int, optional
-    Graph connectivity, at least 2. Default is 16.
-ef_construction : int, optional
-    Construction search breadth. Default is 200.
-nbits : int, optional
-    Total quantization bits, 1 through 9. Default is 8.
-metric : {'l2', 'ip', 'innerproduct'}, optional
-    Squared L2 or inner-product distance. Default is 'l2'. Vectors are not
-    automatically normalized for cosine similarity.
-random_seed : int, optional
-    Graph random seed. Default is 100.
-
-Notes
------
-Build or load the index before searching or saving. Concurrent reads are
-supported; an overlapping mutation raises RuntimeError.)doc"
+            kIndexDoc
         )
         .def(
             "build",
diff --git a/python_bindings/ivf_bindings.cpp b/python_bindings/ivf_bindings.cpp
index 20b81065..0f96da2c 100644
--- a/python_bindings/ivf_bindings.cpp
+++ b/python_bindings/ivf_bindings.cpp
@@ -21,7 +21,6 @@
 #include "rabitqlib/index/ivf/initializer.hpp"
 #include "rabitqlib/index/ivf/ivf.hpp"
 #include "rabitqlib/index/types.hpp"
-#include "rabitqlib/utils/packed.hpp"
 
 namespace py = pybind11;
 
@@ -199,18 +198,7 @@ class IvfIndex {
         if (!built_) {
             throw std::runtime_error("IvfIndex must be built or loaded before remove");
         }
-        std::vector<rabitqlib::PID> to_remove;
-        to_remove.reserve(static_cast<size_t>(ids_array.shape(0)));
-        const rabitqlib::detail::ConstPackedArrayView<int64_t> values(
-            reinterpret_cast<const char*>(ids_array.data())
-        );
-        for (py::ssize_t i = 0; i < ids_array.shape(0); ++i) {
-            const int64_t id = values[static_cast<size_t>(i)];
-            if (id < 0 || static_cast<uint64_t>(id) >= index_->num_stored_points()) {
-                throw std::invalid_argument("ids must be in [0, num_stored_points)");
-            }
-            to_remove.push_back(static_cast<rabitqlib::PID>(id));
-        }
+        auto to_remove = checked_removal_ids(ids_array, index_->num_stored_points());
         py::gil_scoped_release release;
         return index_->remove(to_remove.data(), to_remove.size());
     }
@@ -355,8 +343,8 @@ class IvfIndex {
 void register_ivf(py::module_& m) {
     using namespace rabitqlib::python_bindings;
 
-    py::class_<IvfIndex>(
-        m, "IvfIndex", R"doc(An approximate nearest-neighbor index using IVF.
+    static constexpr char kIndexDoc[] =
+        R"doc(An approximate nearest-neighbor index using IVF.
 
 Parameters
 ----------
@@ -390,8 +378,9 @@ Examples
 >>> index.build(x, x.mean(axis=0, keepdims=True), np.zeros(len(x), dtype=np.uint32))
 >>> ids, distances = index.search(x[:2], k=3, nprobe=1)
 >>> ids.shape, distances.shape
-((2, 3), (2, 3)))doc"
-    )
+((2, 3), (2, 3)))doc";
+
+    py::class_<IvfIndex>(m, "IvfIndex", kIndexDoc)
         .def(
             py::init<
                 size_t,
@@ -406,41 +395,7 @@ Examples
             py::arg("nbits"),
             py::arg("metric") = "l2",
             py::arg("initializer") = "auto",
-            R"doc(An approximate nearest-neighbor index using IVF.
-
-Parameters
-----------
-dim : int
-    Original vector dimension, from 64 through 65536.
-max_elements : int
-    Number of vectors required by the initial build. Later adds grow
-    storage.
-num_clusters : int
-    Number of supplied centroids.
-nbits : int
-    Total quantization bits, 1 through 9, or 32 for owned float32 vectors
-    with one-bit filtering and raw reranking.
-metric : {'l2', 'ip', 'innerproduct'}, optional
-    Squared L2 or inner-product distance. Default is 'l2'. Vectors are not
-    automatically normalized for cosine similarity.
-initializer : {'auto', 'flat', 'flat_rabitq', 'hnsw'}, optional
-    Centroid routing strategy. Default is 'auto'.
-
-Notes
------
-Build or load the index before searching or saving. Concurrent reads are
-supported; an overlapping mutation raises RuntimeError.
-
-Examples
---------
->>> import numpy as np
->>> from rabitqlib import IvfIndex
->>> x = np.random.default_rng(42).normal(size=(80, 64)).astype(np.float32)
->>> index = IvfIndex(64, len(x), num_clusters=1, nbits=4)
->>> index.build(x, x.mean(axis=0, keepdims=True), np.zeros(len(x), dtype=np.uint32))
->>> ids, distances = index.search(x[:2], k=3, nprobe=1)
->>> ids.shape, distances.shape
-((2, 3), (2, 3)))doc"
+            kIndexDoc
         )
         .def(
             "build",
diff --git a/python_bindings/symqg_bindings.cpp b/python_bindings/symqg_bindings.cpp
index 94bcfdc8..7f49e0f1 100644
--- a/python_bindings/symqg_bindings.cpp
+++ b/python_bindings/symqg_bindings.cpp
@@ -219,8 +219,8 @@ class SymqgIndex {
 void register_symqg(py::module_& m) {
     using namespace rabitqlib::python_bindings;
 
-    py::class_<SymqgIndex>(
-        m, "SymqgIndex", R"doc(An approximate nearest-neighbor index using SymphonyQG.
+    static constexpr char kIndexDoc[] =
+        R"doc(An approximate nearest-neighbor index using SymphonyQG.
 
 Parameters
 ----------
@@ -238,33 +238,43 @@ quantization_bits : {0, 4, 8}, optional
 Notes
 -----
 Build or load the index before searching or saving. Concurrent reads are
-supported; an overlapping mutation raises RuntimeError.)doc"
-    )
-        .def(
-            py::init<size_t, size_t, const std::string&, size_t>(),
-            py::arg("dim"),
-            py::arg("max_degree"),
-            py::arg("metric") = "l2",
-            py::arg("quantization_bits") = 0,
-            R"doc(An approximate nearest-neighbor index using SymphonyQG.
+supported; an overlapping mutation raises RuntimeError.)doc";
+
+    static constexpr char kSearchDoc[] = R"doc(Search a batch of query vectors.
 
 Parameters
 ----------
-dim : int
-    Original vector dimension, from 64 through 65536.
-max_degree : int
-    Graph degree bound; a positive multiple of 32.
-metric : {'l2', 'ip', 'innerproduct'}, optional
-    Squared L2 or inner-product distance. Default is 'l2'. Vectors are not
-    automatically normalized for cosine similarity.
-quantization_bits : {0, 4, 8}, optional
-    Centroid-relative RaBitQ storage at 4 or 8 bits, or raw float32 at 0
-    (default).
+queries : array_like of float, shape (n_queries, dim)
+    Vectors in the original, unpadded coordinates. Converted to contiguous
+    float32; a copy may be made.
+k : int
+    Number of neighbors requested, between 1 and num_points.
+ef : int
+    Search breadth. Must be positive and at least k.
+num_threads : int, optional
+    Worker count. Default is 1; 0 selects the available CPU count.
+
+Returns
+-------
+ids : ndarray of uint32, shape (n_queries, k)
+    Point IDs of the selected neighbors.
+distances : ndarray of float32, shape (n_queries, k)
+    Lower values are better; squared L2 estimates for l2 and inner-product
+    distance estimates for ip. These are distances, not similarity scores.
 
 Notes
 -----
-Build or load the index before searching or saving. Concurrent reads are
-supported; an overlapping mutation raises RuntimeError.)doc"
+Returns (ids, distances), in that order. Array shapes and value ranges are
+checked at runtime, not by static type checkers.)doc";
+
+    py::class_<SymqgIndex>(m, "SymqgIndex", kIndexDoc)
+        .def(
+            py::init<size_t, size_t, const std::string&, size_t>(),
+            py::arg("dim"),
+            py::arg("max_degree"),
+            py::arg("metric") = "l2",
+            py::arg("quantization_bits") = 0,
+            kIndexDoc
         )
         .def(
             "build",
@@ -299,32 +309,7 @@ synchronous call.)doc"
             py::arg("k"),
             py::arg("ef"),
             py::arg("num_threads") = 1,
-            R"doc(Search a batch of query vectors.
-
-Parameters
-----------
-queries : array_like of float, shape (n_queries, dim)
-    Vectors in the original, unpadded coordinates. Converted to contiguous
-    float32; a copy may be made.
-k : int
-    Number of neighbors requested, between 1 and num_points.
-ef : int
-    Search breadth. Must be positive and at least k.
-num_threads : int, optional
-    Worker count. Default is 1; 0 selects the available CPU count.
-
-Returns
--------
-ids : ndarray of uint32, shape (n_queries, k)
-    Point IDs of the selected neighbors.
-distances : ndarray of float32, shape (n_queries, k)
-    Lower values are better; squared L2 estimates for l2 and inner-product
-    distance estimates for ip. These are distances, not similarity scores.
-
-Notes
------
-Returns (ids, distances), in that order. Array shapes and value ranges are
-checked at runtime, not by static type checkers.)doc"
+            kSearchDoc
         )
         .def(
             "search_batch",
@@ -333,32 +318,7 @@ checked at runtime, not by static type checkers.)doc"
             py::arg("k"),
             py::arg("ef"),
             py::arg("num_threads") = 1,
-            R"doc(Search a batch of query vectors.
-
-Parameters
-----------
-queries : array_like of float, shape (n_queries, dim)
-    Vectors in the original, unpadded coordinates. Converted to contiguous
-    float32; a copy may be made.
-k : int
-    Number of neighbors requested, between 1 and num_points.
-ef : int
-    Search breadth. Must be positive and at least k.
-num_threads : int, optional
-    Worker count. Default is 1; 0 selects the available CPU count.
-
-Returns
--------
-ids : ndarray of uint32, shape (n_queries, k)
-    Point IDs of the selected neighbors.
-distances : ndarray of float32, shape (n_queries, k)
-    Lower values are better; squared L2 estimates for l2 and inner-product
-    distance estimates for ip. These are distances, not similarity scores.
-
-Notes
------
-Returns (ids, distances), in that order. Array shapes and value ranges are
-checked at runtime, not by static type checkers.)doc"
+            kSearchDoc
         )
         .def(
             "save",
diff --git a/src/index/qg.cpp b/src/index/qg.cpp
index 12920132..518f751f 100644
--- a/src/index/qg.cpp
+++ b/src/index/qg.cpp
@@ -39,6 +39,12 @@
 
 namespace rabitqlib::symqg {
 
+namespace {
+// Native-endian numeric tag; little-endian files contain the bytes "QTIBARGQ".
+constexpr uint64_t kFormatMagic = 0x5147524142495451ULL;
+constexpr uint32_t kFormatVersion = 1;
+}  // namespace
+
 QuantizedQuery::QuantizedQuery(
     const float* rotated_query,
     const float* centroid,
@@ -261,9 +267,6 @@ void QuantizedGraph<float>::save(const char* filename) const {
     }
     output.exceptions(std::ios::badbit | std::ios::failbit);
 
-    // Native-endian numeric tag; little-endian files contain the bytes "QTIBARGQ".
-    constexpr uint64_t kFormatMagic = 0x5147524142495451ULL;
-    constexpr uint32_t kFormatVersion = 1;
     if (quantization_bits_ != 0) {
         output.write(reinterpret_cast<const char*>(&kFormatMagic), sizeof(kFormatMagic));
         output.write(
@@ -333,9 +336,6 @@ void QuantizedGraph<float>::load(const char* filename) {
         read_exact(&value, sizeof(value), field);
     };
 
-    // Native-endian numeric tag; little-endian files contain the bytes "QTIBARGQ".
-    constexpr uint64_t kFormatMagic = 0x5147524142495451ULL;
-    constexpr uint32_t kFormatVersion = 1;
     uint64_t magic = 0;
     read_value(magic, "format marker");
     if (magic == kFormatMagic) {